OpenAI Debuts GPT-Live-1 API for Real-Time Full-Duplex Voice
The new model enables simultaneous listening and speaking, allowing developers to build fluid voice agents that handle interruptions naturally.
OpenAI has released GPT-Live-1 into its API, providing developers with a full-duplex voice model capable of listening and speaking simultaneously. This release marks a shift toward native audio-to-audio processing to eliminate the latency and rigidity of traditional voice AI.
Unlike previous architectures that chained speech-to-text (STT), a large language model (LLM), and text-to-speech (TTS) in a sequential loop, GPT-Live-1 handles audio within a single model. According to OpenAI, this approach simplifies the voice layer and allows the system to respond to interruptions in real-time. In early evaluations for Speak, a language tutor application, the model reduced interruptions by nearly 80% compared to traditional turn-based systems. Performance gains are further evidenced by the Full Duplex Bench, where GPT-Live-1 improved performance by 30 percentage points over GPT-Realtime-2.1.
A Modular Approach to Reasoning
The architecture of GPT-Live-1 is designed to function as a specialized front-end voice layer. Rather than attempting to handle all cognitive tasks internally, the model can delegate complex reasoning and tool calling to more powerful backend models, such as GPT-6 Astra or other third-party alternatives. This separation allows developers to maintain a fluid conversational interface while leveraging heavy-duty processing only when necessary.
Industry Implications
By exposing this capability via API, OpenAI is enabling a new generation of highly fluid voice agents for sectors such as telephony, customer support, and education. The ability to decouple the "voice layer" from the "reasoning layer" allows developers to optimize for both cost and speed. This modularity means that simple conversational fillers can be handled by the low-latency front-end, while complex queries are routed to the backend, potentially making AI voice interactions indistinguishable from human phone calls. OpenAI has set the pricing for the GPT-Live-1 front-end voice layer at $0.05 per minute.
The Path Forward
As developers begin integrating GPT-Live-1, the industry will be watching to see how this full-duplex capability affects user retention in voice-first applications. While the technical benchmarks show significant improvement over GPT-Realtime-2.1, the real-world efficacy of the delegation system—specifically how seamlessly GPT-Live-1 hands off tasks to GPT-6 Astra—remains the primary point of interest for enterprise implementation.